The final part. Parts I and II covered how to talk to Claude and how to think with it. This is the toolbox — which tool to reach for when. It’s the most engineering-heavy part of the series, and it’s where the gold gets made.
Part III: The Toolbox
7. Chat vs. Code vs. Agents vs. Skills vs. Design
Key takeaway: Different tools for different jobs. Use the right one or waste your time.
Claude Chat is for thinking, exploration, and one-off problems where you’re the decision-maker.
Analyzing user interview transcripts
Drafting copy and iterating on it
Thinking through product strategy
Writing and messaging
Any time you’re going to read the output and decide what to do next
Claude Code (in VS Code or the terminal) is for files, automation, and deterministic output.
Processing files (renaming, organizing, transforming data)
Writing scripts that do the same thing repeatedly
Data transformation and analysis
Anything where you need output you can verify and run
Agents — this is where the gold is made — are for autonomous, repeated decision-making over time.
Systems that run on a schedule without human intervention
Claude observes results, adapts, and acts again
Our content agent is the proof: it sweeps the industry, selects stories, and drafts publishable commentary every single day with no one touching it — driving useful content for our customers and improving our SEO
Start hand-managed and earn your way to automation — don’t over-engineer a process you haven’t run manually first!
Skills are the backbone of effective use — and the raw material of agents. A skill is packaged expertise: a written-down playbook Claude loads whenever the task calls for it.
Turn any workflow you repeat into a skill — your weekly update format, your brand voice, your code review checklist, how you structure a customer interview
Define what “good” looks like once, instead of re-explaining it in every conversation
Skills are how agents get trained: an agent without skills is an intern with no onboarding
Start by noticing what you keep re-typing across chats — that’s your first skill
Design is for when the output needs to be seen, not read.
Landing pages, mockups, and prototypes you can actually click through
Pitch decks and one-pagers that look finished, not generated
Exploring three visual directions before committing to one
Turning “here’s roughly what I mean” into something you can put in front of a real user by end of day
8. Use Plan Mode in Claude Code
Key takeaway: Think before you code. Plan mode catches mistakes before they cost you time.
When you open Claude Code, work in plan mode first. Claude thinks through the approach, validates the plan with you, and only then generates code.
Our rule is 90% planning, 10% execution.
This is the most common mistake I see inside Claude Code: too little planning.
Use plan mode by default when the task is:
Complex or multi-step
Modifying existing code (you want Claude to understand dependencies first)
Anything where a wrong first move wastes time
9. Project Instructions Shape Your Default Behavior
Key takeaway: Set up your project instructions once. They persist. Make them work for you instead of explaining the same context over and over.
Project instructions are the system-level context that lives in every conversation in a project. They’re not documentation. They’re behavioral directives.
What goes in project instructions:
Your thinking style (do you want to be challenged? Do you want affirmation? Be explicit.)
Context that applies to most conversations (company background, customer profile, product stage)
Constraints and principles that shape your work
How you want Claude to behave toward you specifically
What doesn’t go in instructions:
One-time information specific to a single conversation (put that in the chat)
Secrets, credentials, or sensitive data (never)
Detailed specs for individual tasks
Example from my own setup:
My thinking style: Challenge my assumptions. Point out weak logic, hidden premises, and unfounded leaps. If you agree with me, explain why with reasoning and counterarguments. No flattery, no “great question.” Direct, neutral, constructive.
My preferences: Short punchy sentences at emotional peaks. Elevated but not pretentious. No jargon.
What I’m optimizing for: Clarity and truthfulness over cleverness.
Now every time you paste something in, Claude knows you’re not looking for affirmation. You’re looking for holes. That changes everything about the feedback you get.
And iterate your instructions. After a week or two, you’ll notice what you keep explaining in every chat. Add it. If Claude keeps doing X when you want Y, make the instruction more explicit. Good instructions evolve.
10. The Most Important Rule
Key takeaway: Claude accelerates learning. It doesn’t replace it.
Everything above makes you faster. This is the rule that makes you better.
If you don’t know Python, don’t use Claude Code to write complex scripts you’ll never understand. Do use Claude Code to write your first script with Claude doing 80% of the work. Over three iterations, you’ll have begun to learned Python and built something useful.
The rule: if you don’t understand the output, ask Claude to explain it in smaller pieces — or go learn the concept separately.
Don’t copy-paste code you don’t understand — it’s a surefire way to make an ass of yourself.
Four interns learned this playbook in six weeks, and the shipped list in Part I is what came out the other side.
We’re already building next summer’s class. Fair warning: this is not a resume-padding internship. There is no coffee run. Interns ship production code with real customers on the other end, and they stand up in a daily standup every morning and account for it. Some kids will read that and run. The right ones will read it twice.
If that’s your kid — or a student you know, 18 to 21, high school senior through college — applications are open: [APPLICATION LINK]